Darkroom4364/genome-firewall-ecoli

Genome Firewall — E. coli antibiotic response models

Per-drug elastic-net logistic regression models predicting, from AMRFinderPlus gene/mutation features of a reconstructed E. coli genome, whether an antibiotic is likely to fail / likely to work / no-call (abstain when evidence is weak).

Research prototype. All predictions must be confirmed with standard laboratory susceptibility testing. Not a medical device.

Metrics (held-out genetic groups — clusters never seen in training)

drug	balanced acc	R recall	S recall	AUROC	no-call rate	acc when called
ciprofloxacin	0.916	0.93	0.91	0.967	0.14	0.944
gentamicin	0.944	0.91	0.98	0.962	0.09	0.969
ampicillin	0.823	0.89	0.75	0.916	0.35	0.908
trimethoprim/sulfamethoxazole	0.946	0.96	0.93	0.972	0.39	0.965
cefotaxime	0.950	0.93	0.97	0.980	0.13	0.964

Evaluation: skani (ANI >= 99.5%) single-linkage clusters; train/cal/test/hidden split by cluster. Labels: BV-BRC lab-measured AST re-derived against EUCAST v16.1 breakpoints.

Contents

• models/<drug>/model.skops — the calibrated classifier (skops format, safe to load)
• models/<drug>/nocall_bands.json — asymmetric class-conditional conformal bands
• features/feature_columns.json — the 600 feature names (AMRFinderPlus 4.2.7 / DB 2026-03-24.1)

Usage

import skops.io as sio
from huggingface_hub import hf_hub_download
path = hf_hub_download(repo_id="Darkroom4364/genome-firewall-ecoli", filename="models/ciprofloxacin/model.skops")
bundle = sio.load(path, trusted=sio.get_untrusted_types(file=path))
clf = bundle["model"]  # calibrated elastic-net LR; bundle also has no-call bands + metadata
# X = feature vector aligned with features/feature_columns.json (0/1 per AMR element)
# p_fail = clf.predict_proba(X)[0][1]

Pipeline to produce features from a FASTA + full report: https://github.com/Trista1208/The-Genome-Reader (branch sprint/baseline)

Built at Hack-Nation 6th Global AI Hackathon.